Personal work twin: compiles versioned persona prompts from mema preference memories and injects them into plan workflows.
Single tool 'twin' with severe definition gaps. Tool name lacks action verb (should be verb_noun like 'compile_persona' or 'manage_twin'). Description is 400+ chars in Chinese, exceeds 1024-char guideline, and lacks English translation for non-Chinese speakers. Input schema present but parameters lack individual descriptions, 'action' has a description but 'data' is generic ('动作参数字典,具体字段因动作而异' = 'action parameter dict, specific fields vary by action'). No output schema documented. No enum constraint on 'action' parameter despite 20+ known values. Error handling returns generic dicts without recovery guidance. Tool is a dispatcher masquerading as a single tool, should be split into 20+ focused tools (write_preference, get_preference, compile_persona, submit_version, etc.). No parameter validation rules, no field-level descriptions for 'data' object.
个人分身 twin:按工作性质沉淀用户工作偏好,编译版本化 persona prompt,经交付任务流注入执行,并提供定时任务建议。 动作:write / get / compile / submit / rollback / status / taxonomy / pending / resolve / void / task_start / task_submit / task_resume / task_revise / task_close / task_recent / task_get / todo / help。 先 twin(action="help") 查看各动作参数与引导。compile 返回素材包,由当前会话模型编译(建议在强模型会话中执行),submit 提交回库落版本并写文件镜像。
Tool name 'twin' lacks action verb. Should be verb_noun (e.g., 'compile_persona', 'manage_twin_preference'). LLMs cannot infer intent from a bare noun.
Input parameter 'data' has no type schema or field descriptions. Marked as 'object' with generic description '动作参数字典,具体字段因动作而异'. LLMs cannot validate or construct valid payloads without per-field schemas.
Parameter 'action' accepts 20+ values but no enum constraint. Free-form string invites hallucinated actions. Must declare enum: ['write', 'get', 'compile', 'submit', 'rollback', 'status', 'taxonomy', 'pending', 'resolve', 'void', 'task_start', 'task_submit', 'task_resume', 'task_revise', 'task_close', 'task_recent', 'task_get', 'todo', 'help'].
No output schema documented. Responses are unstructured dicts with variable fields (ok, error, reason, actions, mema_notices, etc.). LLMs cannot plan downstream calls or extract structured data.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | 2026-07-28+ | v2 |
Single tool dispatches 20+ distinct operations (write, get, compile, submit, task_start, etc.). Violates single-responsibility principle. Should be split into focused tools: write_preference, get_preference, compile_persona, submit_version, start_task, etc. Current design forces LLM to reason about action strings instead of tool selection.